|
on Econometrics |
| By: | Fu Ouyang; Thomas Tao Yang |
| Abstract: | We propose a new approach to the semiparametric analysis of panel data binary choice models with fixed effects and dynamics (lagged dependent variables). The model we consider has the same random utility framework as in Honore and Kyriazidou ´ (2000). We demonstrate that, with additional serial dependence conditions on the process of deterministic utility and tail restrictions on the error distribution, the (point) identification of the model can proceed in two steps, and only requires matching the value of an index function of explanatory variables over time, as opposed to that of each explanatory variable. Our identification approach motivates an easily implementable, two-step maximum score (2SMS) procedure – producing estimators whose rates of convergence, in contrast to Honore and Kyriazidou ´ ’s (2000) methods, are independent of the model dimension. We then derive the asymptotic properties of the 2SMS procedure and propose bootstrap-based distributional approximations for inference. Monte Carlo evidence indicates that our procedure performs adequately in finite samples. We then apply the proposed estimators to study labor market dependence and the effects of health shocks, using data from the Household, Income and Labor Dynamics in Australia (HILDA) survey. |
| Keywords: | Semiparametric estimation; Binary choice model; Panel data; Fixed effects; Dynamics; Maximum score; Bootstrap |
| JEL: | C14 C23 C35 |
| Date: | 2020–05 |
| URL: | http://d.repec.org/n?u=RePEc:acb:cbeeco:2020-671&r=all |
| By: | António Rua; Francisco Dias; Maximiano Pinheiro |
| Abstract: | Along with the advances of statistical data collection worldwide, dynamic factor models have gained prominence in economics and finance when dealing with data rich environments. Although factor models have been typically applied to two-dimensional data, three-way array data sets are becoming increasingly available. Motivated by the tensor decomposition literature, we propose a dynamic factor model for three-way data. We show that this modeling strategy is flexible while remaining quite parsimonious, in sharp contrast with previous approaches. We discuss identification and put forward a set of identifying restrictions that enhance the interpretation of the model. We propose an estimation procedure based on maximum likelihood using the Expectation-Conditional Maximization algorithm and assess the finite sample properties of the estimator through a Monte Carlo study. In the empirical application, we apply the model to inflation data for nineteen euro area countries and fifty-five products covering the last two decades. |
| JEL: | C38 C51 E31 |
| Date: | 2020 |
| URL: | http://d.repec.org/n?u=RePEc:ptu:wpaper:w202007&r=all |
| By: | Tobias Hartl; Rolf Tschernig; Enzo Weber |
| Abstract: | We develop a generalization of correlated trend-cycle decompositions that avoids prior assumptions about the long-run dynamic characteristics by modelling the permanent component as a fractionally integrated process and incorporating a fractional lag operator into the autoregressive polynomial of the cyclical component. We relate the model to the Beveridge-Nelson decomposition and derive a modified Kalman filter estimator for the fractional components. Identification and consistency of the maximum likelihood estimator are shown. For US macroeconomic data we demonstrate that, unlike non-fractional correlated unobserved components models, the new model estimates a smooth trend together with a cycle hitting all NBER recessions. |
| Date: | 2020–05 |
| URL: | http://d.repec.org/n?u=RePEc:arx:papers:2005.05266&r=all |
| By: | Wang, Dandan; Escribano Saez, Alvaro |
| Abstract: | The use of machine learning (ML) models has been shown to have advantages over alternative and more traditional time series models in the presence of big data. One of the most successful ML forecasting procedures is the Random Forest (RF) machine learning algorithm. In this paper we propose a mixed RF approach for modeling departures from linearity, instead of starting with a completely nonlinear or nonparametric model. The methodology is applied to the weekly forecasts of gasoline prices that are cointegrated with international oil prices and exchange rates. The question of interest is whether gasoline prices react asymmetrically to increases in oil prices rather than to decreases in oil prices, the "rockets and feathers" hypothesis. In this literature most authors estimate parametric nonlinear error correction models using nonlinear least squares. Recent specifications for nonlinear error correction models include threshold autoregressive models (TAR), double threshold error correction models (ECM) or double threshold smooth transition autoregressive (STAR) models. In this paper, we describe the econometric methodology that combines linear dynamic autoregressive distributed lag (ARDL) models with cointegrated variables with added nonlinear components, or price asymmetries, estimated by the powerful tool of RF. We apply our mixed RF specification strategy to weekly prices of the Spanish gasoline market from 2010 to 2019. We show that the new mixed RF error correction model has important advantages over competing parametric and nonparametric models, in terms of the generality of model specification, estimation and forecasting. |
| Keywords: | Mixed Random Forest; Random Forest; Machine Learning; Nonlinear Error Correction; Cointegration; Rockets And Feathers Hypothesis; Forecasting Gasoline Prices |
| JEL: | L71 L13 D43 C53 C52 C24 B23 |
| Date: | 2020–06–04 |
| URL: | http://d.repec.org/n?u=RePEc:cte:werepe:30557&r=all |
| By: | Laurent Pauwels; Peter Radchenko; Andrey L. Vasnev |
| Abstract: | The majority of financial data exhibit asymmetry and heavy tails, which makes forecasting the entire density critically important. Recently, a forecast combination methodology has been developed to combine predictive densities. We show that combining individual predictive densities that are skewed and/or heavy-tailed results in significantly reduced skewness and kurtosis. We propose a solution to overcome this problem by deriving optimal log score weights under Higher-order Moment Constraints (HMC). The statistical properties of these weights, such as consistency and asymptotic distribution, are investigated theoretically and through a simulation study. An empirical application that uses the S&P 500 daily index returns illustrates that the proposed HMC weight density combinations perform very well relative to other combination methods. |
| Keywords: | Forecast combinations, Predictive densities, Moment constraints, Financial data. |
| JEL: | C53 C58 |
| Date: | 2020–05 |
| URL: | http://d.repec.org/n?u=RePEc:een:camaaa:2020-45&r=all |
| By: | Felix Elwert; Elan Segarra |
| Abstract: | Instrumental variables (IV) estimation suffers selection bias when the analysis conditions on the treatment. Judea Pearl's early graphical definition of instrumental variables explicitly prohibited conditioning on the treatment. Nonetheless, the practice remains common. In this paper, we derive exact analytic expressions for IV selection bias across a range of data-generating models, and for various selection-inducing procedures. We present four sets of results for linear models. First, IV selection bias depends on the conditioning procedure (covariate adjustment vs. sample truncation). Second, IV selection bias due to covariate adjustment is the limiting case of IV selection bias due to sample truncation. Third, in certain models, the IV and OLS estimators under selection bound the true causal effect in large samples. Fourth, we characterize situations where IV remains preferred to OLS despite selection on the treatment. These results broaden the notion of IV selection bias beyond sample truncation, replace prior simulation findings with exact analytic formulas, and enable formal sensitivity analyses. |
| Date: | 2020–05 |
| URL: | http://d.repec.org/n?u=RePEc:arx:papers:2005.09583&r=all |
| By: | Taras Bodnar; Solomiia Dmytriv; Yarema Okhrin; Nestor Parolya; Wolfgang Schmid |
| Abstract: | In this paper, using the shrinkage-based approach for portfolio weights and modern results from random matrix theory we construct an effective procedure for testing the efficiency of the expected utility (EU) portfolio and discuss the asymptotic behavior of the proposed test statistic under the high-dimensional asymptotic regime, namely when the number of assets $p$ increases at the same rate as the sample size $n$ such that their ratio $p/n$ approaches a positive constant $c\in(0,1)$ as $n\to\infty$. We provide an extensive simulation study where the power function and receiver operating characteristic curves of the test are analyzed. In the empirical study, the methodology is applied to the returns of S\&P 500 constituents. |
| Date: | 2020–05 |
| URL: | http://d.repec.org/n?u=RePEc:arx:papers:2005.04761&r=all |
| By: | Michael Creel |
| Abstract: | This paper deals with Laplace type methods used with moment-based, simulation-based, econometric estimators. It shows that confidence intervals based upon quantiles of a tuned MCMC chain may have coverage which is far from the nominal level. It discusses how neural networks may be used to easily and automatically reduce the dimension of an initial set of moments to the minimum number of moments needed to maintain identification. When estimation and inference is based on the neural moments, which are the result of filtering moments through a trained neural net, confidence intervals have correct coverage in almost all cases, and departures from correct coverage are small. |
| Keywords: | neural networks, Laplace type estimators, simulation-based estimation |
| JEL: | C11 C12 C13 C45 |
| Date: | 2020–06 |
| URL: | http://d.repec.org/n?u=RePEc:bge:wpaper:1182&r=all |
| By: | Tobias Hartl |
| Abstract: | We combine high-dimensional factor models with fractional integration methods and derive models where nonstationary, potentially cointegrated data of different persistence is modelled as a function of common fractionally integrated factors. A two-stage estimator, that combines principal components and the Kalman filter, is proposed. The forecast performance is studied for a high-dimensional US macroeconomic data set, where we find that benefits from the fractional factor models can be substantial, as they outperform univariate autoregressions, principal components, and the factor-augmented error-correction model. |
| Date: | 2020–05 |
| URL: | http://d.repec.org/n?u=RePEc:arx:papers:2005.04897&r=all |
| By: | Yasumasa Matsuda; Xin Yuan |
| Abstract: | This paper conducts a multivariate extension of isotropic Levy- driven CARMA random fileds on Rd proposed by Brockwell and Matsuda (2017). Univariate CARMA models are defined as moving averages of a Levy sheet with CARMA kernels defined by AR and MA polynomials. We define multivariate CARMA models by a multivariate extension of CARMA kernels with matrix valued AR and MA polynomials. For the multivariate CARMA models, we derive the spectral density functions as explicit parametric func- tions. Given multivariate irregularly spaced data on R2, we propose Whittle estimation of CARMA parameters to minimize Whittle likelihood given with periodogram matrices and clarify conditions under which consistency and as- ymptotic normality hold under the so called mixed asymptotics. We nally in- troduce a method to conduct kriging for irregularly spaced data on R2 by mul- tivariate CARMA random fields with the estimated parameters in a Bayesian way and demonstrate the empirical properties by tri-variate spatial dataset of simulation and of US precipitation data. |
| Date: | 2020–04 |
| URL: | http://d.repec.org/n?u=RePEc:toh:dssraa:113&r=all |
| By: | Bo E. Honor\'e; Martin Weidner |
| Abstract: | This paper presents new moment conditions for dynamic panel data logit models with fixed effects. After introducing the new moment conditions we explore their implications for identification and estimation of the model parameters that are common to all individuals, and we find that those common model parameters are estimable at root-n rate for many more dynamic panel logit models than was previously known in the literature. A GMM estimator that is based on the new moment conditions is shown to perform well in Monte Carlo simulations and in an empirical illustration to labor force participation. |
| Date: | 2020–05 |
| URL: | http://d.repec.org/n?u=RePEc:arx:papers:2005.05942&r=all |
| By: | Helmut Lütkepohl |
| Abstract: | In conventional structural vector autoregressive (VAR) models it is assumed that there are at most as many structural shocks as there are variables in the model. It is pointed out that heteroskedasticity can be used to identify more shocks than variables. However, even if there is heteroskedasticity, the number of shocks that can be identified is limited. A number of results are provided that allow a researcher to assess how many shocks can be identified from specific forms of heteroskedasticity. |
| Keywords: | Structural vector autoregression, identification through heteroskedasticity, structural shocks |
| JEL: | C32 |
| Date: | 2020 |
| URL: | http://d.repec.org/n?u=RePEc:diw:diwwpp:dp1871&r=all |
| By: | Qiu, Yue (Shanghai University of International Business and Economics); Xie, Tian (Shanghai University of Finance and Economics); Yu, Jun (School of Economics, Singapore Management University) |
| Abstract: | This paper introduces novel methods to combine forecasts made by machine learning techniques. Machine learning methods have found many successful applications in predicting the response variable. However, they ignore model uncertainty when the relationship between the response variable and the predictors is nonlinear. To further improve the forecasting performance, we propose a general framework to combine multiple forecasts from machine learning techniques. Simulation studies show that the proposed machine-learning-based forecast combinations work well. In empirical applications to forecast key macroeconomic and financial variables, we find that the proposed methods can produce more accurate forecasts than individual machine learning techniques and the simple average method, later of which is known as hard to beat in the literature. |
| Keywords: | Model uncertainty; Machine learning; Nonlinearity; Forecast combinations |
| JEL: | C52 C53 |
| Date: | 2020–05–11 |
| URL: | http://d.repec.org/n?u=RePEc:ris:smuesw:2020_013&r=all |
| By: | Canepa, Alessandra (University of Turin) |
| Abstract: | A Bartlett corrected likelihood ratio test for linear restrictions on the cointegrating relations is examined in Johansen (2000). Simulation results show that the performance of the corrected LR test statistic is highly dependent on the values of the parameters of the model. In order to reduce this dependency, it is proposed that the ?nite sample expectation of the LR test be estimated using the bootstrap. It is found that the bootstrap Bartlett correction often succeeds in this task. |
| Date: | 2020–03 |
| URL: | http://d.repec.org/n?u=RePEc:uto:dipeco:202007&r=all |
| By: | Juan Carlos Escanciano |
| Abstract: | One of the most important empirical findings in microeconometrics is the pervasiveness of heterogeneity in economic behaviour (cf. Heckman 2001). This paper shows that cumulative distribution functions and quantiles of the nonparametric unobserved heterogeneity have an infinite efficiency bound in many structural economic models of interest. The paper presents a relatively simple check of this fact. The usefulness of the theory is demonstrated with several relevant examples in economics, including, among others, the proportion of individuals with severe long term unemployment duration, the average marginal effect and the proportion of individuals with a positive marginal effect in a correlated random coefficient model with heterogenous first-stage effects, and the distribution and quantiles of random coefficients in linear, binary and the Mixed Logit models. Monte Carlo simulations illustrate the finite sample implications of our findings for the distribution and quantiles of the random coefficients in the Mixed Logit model. |
| Date: | 2020–05 |
| URL: | http://d.repec.org/n?u=RePEc:arx:papers:2005.08611&r=all |
| By: | Fu Ouyang; Thomas Tao Yang; Hanghui Zhang |
| Abstract: | We study (point) identification of preference coefficients in semiparametric discrete choice models for bundles. The approach to the identification uses an “identification at infinity†(Chamberlain (1986)) insight in combination with median independence restrictions on unobservables. We propose two-stage maximum score (MS) estimators and show their consistency |
| Keywords: | Bundle choices, semiparametric model, median independence, identification at infinity, maximum score estimation. |
| JEL: | C14 C23 C35 |
| Date: | 2020–05 |
| URL: | http://d.repec.org/n?u=RePEc:acb:cbeeco:2020-672&r=all |
| By: | Enrique Martínez-García |
| Abstract: | This paper considers the characterization of the reduced-form solution of a large class of linear rational expectations models. I show that under certain conditions, if a solution exists and is unique, it can be cast in finite-order VAR form. I also investigate the conditions for the VAR form to be stationary with a well-defined residual variance-covariance matrix in equilibrium, for the shocks to be recoverable, and for local identification of the structural parameters for estimation from the sample likelihood. An application to the workhorse New Keynesian model with accompanying Matlab codes illustrates the practical use of the finite-order VAR representation. In particular, I argue that the identification of monetary policy shocks based on structural VARs can be more closely aligned with theory using the finite-order VAR form of the model solution characterized in this paper. |
| Keywords: | Linear Rational Expectations Models; Finite-Order Vector Autoregressive Representation; Sylvester Matrix Equation; New Keynesian Model; Monetary Policy Shocks |
| JEL: | C32 C62 C63 E37 |
| Date: | 2020–05–29 |
| URL: | http://d.repec.org/n?u=RePEc:fip:feddgw:88096&r=all |
| By: | Sanghyun Hong (University of Canterbury); W. Robert Reed (University of Canterbury) |
| Abstract: | The purpose of this study is to show how Monte Carlo analysis of meta-analytic estimators can be used to select estimators for specific research situations. Our analysis conducts 1,620 individual experiments, where each experiment is defined by a unique combination of sample size, effect heterogeneity, effect size, publication selection mechanism, and other research characteristics. We compare eleven estimators commonly used in medicine, psychology, and the social sciences. These are evaluated on the basis of bias, mean squared error (MSE), and coverage rates. For our experimental design, we reproduce simulation environments from four recent studies: Stanley, Doucouliagos, & Ioannidis (2017), Alinaghi & Reed (2018), Bom & Rachinger (2019), and Carter et al. (2019a). We demonstrate that relative estimator performance differs across performance measures. An estimator that may be especially good with respect to MSE may perform relatively poorly with respect to coverage rates. We also show that sample size and effect heterogeneity are important determinants of relative estimator performance. We use these results to demonstrate how the observable characteristics of sample size and effect heterogeneity can guide the meta-analyst in choosing the estimators most appropriate for their research circumstances. All of the programming code and output files associated with this project are available at https://osf.io/pr4mb/. |
| Keywords: | Meta-analysis, Estimator performance, Publication bias, Simulation design, Monte Carlo, Experiments |
| JEL: | B41 C15 C18 |
| Date: | 2020–06–01 |
| URL: | http://d.repec.org/n?u=RePEc:cbt:econwp:20/10&r=all |
| By: | Masato Hisakado; Shintaro Mori |
| Abstract: | We discuss the parameter estimation of the probability of default (PD), the correlation between the obligors, and a phase transition. In our previous work, we studied the problem using the beta-binomial distribution. A non-equilibrium phase transition with an order parameter occurs when the temporal correlation decays by power law. In this article, we adopt the Merton model, which uses an asset correlation as the default correlation, and find that a phase transition occurs when the temporal correlation decays by power law. When the power index is less than one, the PD estimator converges slowly. Thus, it is difficult to estimate PD with limited historical data. Conversely, when the power index is greater than one, the convergence speed is inversely proportional to the number of samples. We investigate the empirical default data history of several rating agencies. The estimated power index is in the slow convergence range when we use long history data. This suggests that PD could have a long memory and that it is difficult to estimate parameters due to slow convergence. |
| Date: | 2020–05 |
| URL: | http://d.repec.org/n?u=RePEc:arx:papers:2005.07967&r=all |
| By: | Samuel Gingras; William J. McCausland |
| Abstract: | We introduce a new stochastic duration model for transaction times in asset markets. We argue that widely accepted rules for aggregating seemingly related trades mislead inference pertaining to durations between unrelated trades: while any two trades executed in the same second are probably related, it is extremely unlikely that all such pairs of trades are, in a typical sample. By placing uncertainty about which trades are related within our model, we improve inference for the distribution of durations between unrelated trades, especially near zero. We introduce a normalized conditional distribution for durations between unrelated trades that is both flexible and amenable to shrinkage towards an exponential distribution, which we argue is an appropriate first-order model. Thanks to highly efficient draws of state variables, numerical efficiency of posterior simulation is much higher than in previous studies. In an empirical application, we find that the conditional hazard function for durations between unrelated trades varies much less than what most studies find. We claim that this is because we avoid statistical artifacts that arise from deterministic trade-aggregation rules and unsuitable parametric distributions. |
| Date: | 2020–05 |
| URL: | http://d.repec.org/n?u=RePEc:arx:papers:2005.09166&r=all |
| By: | Christian Bongiorno; Damien Challet |
| Abstract: | We introduce a $k$-fold boosted version of our Boostrapped Average Hierarchical Clustering cleaning procedure for correlation and covariance matrices. We then apply this method to global minimum variance portfolios for various values of $k$ and compare their performance with other state-of-the-art methods. Generally, we find that our method yields better Sharpe ratios after transaction costs than competing filtering methods, despite requiring a larger turnover. |
| Date: | 2020–05 |
| URL: | http://d.repec.org/n?u=RePEc:arx:papers:2005.08703&r=all |
| By: | Håvard Hungnes (Statistics Norway) |
| Abstract: | The paper derives a test for equal predictability of multi-step-ahead system forecasts that is invariant to linear transformations. The test is a multivariate version of the Diebold-Mariano test. An invariant metric for multi-step-ahead system forecasts is necessary as the conclusions otherwise can depend on how the forecasts are reported (e.g., as in levels or differences; or log-levels or growth rates). The test is used in comparing quarterly multi-step-ahead system forecasts made by Statistics Norway with similar forecasts made by Norges Bank. |
| Keywords: | Macroeconomic forecasts; Econometric models; Forecast performance; Forecast evaluation; Forecast comparison |
| JEL: | C32 C53 |
| Date: | 2020–05 |
| URL: | http://d.repec.org/n?u=RePEc:ssb:dispap:931&r=all |
| By: | Massimiliano Mazzanti (University of Ferrara; SEEDS, Italy); Antonio Musolesi (University of Ferrara; SEEDS, Italy) |
| Abstract: | Innovation is a primary engine of sustainable growth. This paper provides new semiparametric econometric policy evaluation methods and estimates a green knowledge production function for a large, 30-year panel dataset of high-income countries. Because of the high degree of uncertainty surrounding the data-generating process and the likely presence of nonlinearities and latent common factors, the paper considers semiparametric panel specifications that extend the parametric multifactor error model and the random trend model. It also adopts a recently proposed information criterion for smooth model selection to compare these semiparametric models and their parametric counterparts. The results indicate that (1) the semiparametric additive specification with individual time trends is the preferred model, (2) threshold effects and nonlinearities are relevant features of the data that are obscured in parametric specifications, and (3) the effect of environmental policy is significant and clearly heterogeneous when modeled as a nonparametric function of certain knowledge inputs. The evidence shows a relevant nonlinear policy inducement effect occurring through R&D investments. |
| Keywords: | Innovation, knowledge, environmental policy, policy assessment, policy heterogeneity, large panels, cross-sectional dependence, factor models, random trend model, spline functions, model selection. |
| Date: | 2020–06 |
| URL: | http://d.repec.org/n?u=RePEc:srt:wpaper:0920&r=all |
| By: | Andrii Babii; Eric Ghysels; Jonas Striaukas |
| Abstract: | This paper introduces structured machine learning regressions for high-dimensional time series data potentially sampled at different frequencies. The sparse-group LASSO estimator can take advantage of such time series data structures and outperforms the unstructured LASSO. We establish oracle inequalities for the sparse-group LASSO estimator within a framework that allows for the mixing processes and recognizes that the financial and the macroeconomic data may have heavier than exponential tails. An empirical application to nowcasting US GDP growth indicates that the estimator performs favorably compared to other alternatives and that the text data can be a useful addition to more traditional numerical data. |
| Date: | 2020–05 |
| URL: | http://d.repec.org/n?u=RePEc:arx:papers:2005.14057&r=all |
| By: | Gianluca Zeni; Matteo Fontana; Simone Vantini |
| Abstract: | In this work we provide a review of basic ideas and novel developments about Conformal Prediction -- an innovative distribution-free, non-parametric forecasting method, based on minimal assumptions -- that is able to yield in a very straightforward way predictions sets that are valid in a statistical sense also in in the finite sample case. The in-depth discussion provided in the paper covers the theoretical underpinnings of Conformal Prediction, and then proceeds to list the more advanced developments and adaptations of the original idea. |
| Date: | 2020–05 |
| URL: | http://d.repec.org/n?u=RePEc:arx:papers:2005.07972&r=all |
| By: | Paolo Gelain; Simone Manganelli |
| Abstract: | We consider two approaches to incorporate judgment into DSGE models. First, Bayesian estimation indirectly imposes judgment via priors on model parameters, which are then mapped into a judgmental interest rate decision. Standard priors are shown to be associated with highly unrealistic judgmental decisions. Second, judgmental interest rate decisions are directly provided by the decision maker and incorporated into a formal statistical decision rule using frequentist procedures. When the observed interest rates are interpreted as judgmental decisions, they are found to be consistent with DSGE models for long stretches of time, but excessively tight in the 1980s and late 1990s and excessively loose in the late 1970s and early 2000s. |
| JEL: | E50 E58 E47 C12 |
| Date: | 2020–05–21 |
| URL: | http://d.repec.org/n?u=RePEc:fip:fedcwq:88033&r=all |
| By: | Yang Feng; Qingfeng Liu |
| Abstract: | We study the nested model averaging method on the solution path for a high-dimensional linear regression problem. In particular, we propose to combine model averaging with regularized estimators (e.g., lasso and SLOPE) on the solution path for high-dimensional linear regression. In simulation studies, we first conduct a systematic investigation on the impact of predictor ordering on the behavior of nested model averaging, then show that nested model averaging with lasso and SLOPE compares favorably with other competing methods, including the infeasible lasso and SLOPE with the tuning parameter optimally selected. A real data analysis on predicting the per capita violent crime in the United States shows an outstanding performance of the nested model averaging with lasso. |
| Date: | 2020–05 |
| URL: | http://d.repec.org/n?u=RePEc:arx:papers:2005.08057&r=all |
| By: | Kalamara, Eleni (King’s College London); Turrell, Arthur (Bank of England); Redl, Chris (International Monetary Fund); Kapetanios, George (King’s College London); Kapadia, Sujit (European Central Bank) |
| Abstract: | We consider the best way to extract timely signals from newspaper text and use them to forecast macroeconomic variables using three popular UK newspapers that collectively represent UK newspaper readership in terms of political perspective and editorial style. We find that newspaper text can improve economic forecasts both in absolute and marginal terms. We introduce a powerful new method of incorporating text information in forecasts that combines counts of terms with supervised machine learning techniques. This method improves forecasts of macroeconomic variables including GDP, inflation, and unemployment, including relative to existing text-based methods. Forecast improvements occur when it matters most, during stressed periods. |
| Keywords: | Text; forecasting; machine learning |
| JEL: | C55 J42 |
| Date: | 2020–05–22 |
| URL: | http://d.repec.org/n?u=RePEc:boe:boeewp:0865&r=all |
| By: | Michael Puglia; Adam Tucker |
| Abstract: | We use machine learning methods to examine the power of Treasury term spreads and other financial market and macroeconomic variables to forecast US recessions, vis-à-vis probit regression. In particular we propose a novel strategy for conducting cross-validation on classifiers trained with macro/financial panel data of low frequency and compare the results to those obtained from standard k-folds cross-validation. Consistent with the existing literature we find that, in the time series setting, forecast accuracy estimates derived from k-folds are biased optimistically, and cross-validation strategies which eliminate data "peeking" produce lower, and perhaps more realistic, estimates of forecast accuracy. More strikingly, we also document rank reversal of probit, Random Forest, XGBoost, LightGBM, neural network and support-vector machine classifier forecast performance over the two cross-validation methodologies. That is, while a k-folds cross-validation indicates tha t the forecast accuracy of tree methods dominates that of neural networks, which in turn dominates that of probit regression, the more conservative cross-validation strategy we propose indicates the exact opposite, and that probit regression should be preferred over machine learning methods, at least in the context of the present problem. This latter result stands in contrast to a growing body of literature demonstrating that machine learning methods outperform many alternative classification algorithms and we discuss some possible reasons for our result. We also discuss techniques for conducting statistical inference on machine learning classifiers using Cochrane's Q and McNemar's tests; and use the SHapley Additive exPlanations (SHAP) framework to decompose US recession forecasts and analyze feature importance across business cycles. |
| Keywords: | Shapley; Probit; XGBoost; Treasury yield curve; Neural network; LightGBM; Recession; Tree ensemble; Support-vector machine; Random forest |
| JEL: | C45 C53 E37 |
| Date: | 2020–05–20 |
| URL: | http://d.repec.org/n?u=RePEc:fip:fedgfe:2020-38&r=all |